CRM AI & Automation
CRM AI and automation, inside the system you already run.
We deploy AI agents, inquiry triage and automated CRM analysis inside Salesforce, HubSpot and Odoo. The model proposes, your team approves, and every action is written to an audit log — on data that has been checked first.
Five capabilities, three platforms, and no second system for your team to learn.
Why AI in a CRM stalls
Four reasons the pilot never becomes the process
Many failed CRM AI projects were technically fine. The model answered, the demo worked, and six months later nobody used it. The causes are the same every time, and none of them are the model.
The process was never defined
The agent was asked to handle "customer enquiries" when the business has eleven kinds, four owners and no agreed definition of resolved. Automating an undefined process produces the same mess, faster.
The data under it cannot be trusted
Duplicate accounts, empty owners, three spellings of the same company, and a stage field nobody updates. The agent retrieves the wrong record and confidently answers from it. No prompt fixes that.
It sits beside the CRM, not inside it
A separate chatbot tool with its own memory cannot see the opportunity, the invoice or the last call. The customer has to repeat themselves, and the team has to copy the outcome back by hand.
Nobody defined what it may not do
No approval step, no escalation rule, no log. The first confident wrong answer goes to a customer, trust collapses, and the agent is switched off — usually permanently.
The order of work is the whole method: process, data, permissions, then model. We will not start with the model. If the first three are not in place, the honest recommendation is a diagnostic or a data cleanup before any AI budget is spent.
What we deploy
Five capabilities, each with a boundary
These are the five things we build. Each one has a named scope, a place it writes to, and a rule about what it is not allowed to do without a person.
Conversational agents that can see the account
An agent on your website, WhatsApp or phone line that answers from your approved documents, checks the CRM record, qualifies the enquiry, books the meeting and writes the transcript back to the right owner. Not a decision tree with a language model bolted on: a scoped agent with named tools, defined data access and an escalation path.
Writes to: contact, lead or case — summary, intent, qualification fields, transcript link, owner. Never: sends an outbound message or changes an owner without approval.
Inquiry triage: classify, summarise, draft
Every inbound message classified by intent and urgency, summarised onto the record, and a reply drafted for a person to send. The support queue stops being a reading task and becomes an approval task. Works on email, forms, chat and ticket channels.
Writes to: case fields, priority, queue, draft response. Never: closes a case or sends a reply on its own.
CRM analysis that arrives before the meeting
Pipeline risk, next-best action, churn signals and forecast variance generated from your own CRM data and delivered where the team already works: a morning digest, a message in the deal channel, a field on the record. Not a BI programme with a twelve-week wait.
Writes to: insight fields, tasks, digests. Never: edits an amount, a stage or a close date.
Native AI, switched on properly
Agentforce, Breeze and Odoo AI arrive with the licence. Many companies own the capability and use almost none of it, because nobody configured the agents, connected the data or tested them against real cases. We activate what you already pay for before proposing anything custom.
Writes to: wherever the native agent is permitted. Never: given broader permissions than the team it serves.
AI where the CRM ends
An agent is only as useful as the systems it can reach. We connect ERP, billing, telephony, e-commerce and the warehouse, so an answer can include stock, invoice status and delivery date instead of just the last email. Integration work with a retry policy, not a one-way script.
Writes to: CRM records from external systems, and back. Never: leaves a failed call silent — every failure has a queue and an owner.
How it runs
Five phases, one workflow at a time
AI deployments fail when they start with a platform-wide ambition. We start with one workflow, prove it against a baseline, and only then widen the scope. Durations below are typical for a single workflow with clean data.
| Phase | Typical duration | Your involvement | You receive |
|---|---|---|---|
| 01 · Audit | 1–2 weeks | 2–3 h per team lead | Opportunity map ranked by volume, cost and risk; data readiness score |
| 02 · Design | 1 week | 1 h review | Agent spec: scope, tools, data access, approval points, escalation |
| 03 · Build and test | 2–4 weeks | Weekly 45-minute review | Working agent in a sandbox, plus an evaluation set of real past cases |
| 04 · Pilot | 2–4 weeks | One team, daily for week one | Accuracy, deflection and handling-time measures against the baseline |
| 05 · Operate | Ongoing | Monthly review | Monitoring, regression tests, prompt and tool updates, monthly report |
The audit is the part that decides whether the project is worth doing. If the workflow is low volume, poorly defined or sitting on unreliable data, we will say so and recommend a smaller target — or none. A pilot that should not have started is more expensive than a pilot that was never sold.
What each phase produces
- Workflow inventory ranked by hours and error cost
- Data readiness score with named blockers
- Agent specification before a line is built
- Evaluation set built from your real historical cases
- Accuracy and escalation thresholds agreed in advance
- Permission model scoped to the team, not the org
- Full audit log of every AI action
- Pilot results measured against a pre-agreed baseline
- Documented prompts, tools and escalation rules
- Kill switch, tested before go-live
Guardrails
The controls decide whether it survives customers
AI in a CRM is a trust problem before it is a technology problem. One wrong email and the team turns the agent off for good. These controls are not optional extras added at the end; they are the deployment.
The rule we do not bend
Nothing writes to the CRM without a rule.
Every agent we deploy has an explicit boundary: what it may read, what it may draft, what it may write, and when it must stop and ask a person. That boundary is written down, tested, and enforced by permissions rather than by prompt wording.
- Human approval for outbound messages and record changes above an agreed threshold
- Every action logged: input, output, model, user, timestamp, record
- Answers grounded in your documents, with the source shown
- Personal data masked before inference where the platform supports it
- AI disclosure on every customer-facing conversation
- A kill switch per agent, tested before go-live
On the EU AI Act: customer-facing conversational AI carries transparency obligations, including telling people they are interacting with a machine. We build that disclosure in by default rather than treating it as a legal afterthought. Where a workflow touches regulated decisions, we will say that it is out of scope for an autonomous agent.
Across the platforms
What "AI in the CRM" means on each platform
All three now ship AI. What it can do, how much autonomy it is given, and where the governance sits are genuinely different — and so is the amount we have to build ourselves.
Most autonomy, most governance
Salesforce · Agentforce
Configurable agents that act across sales, service and marketing, grounded in Data Cloud. The strongest control layer of the three: masking, zero retention, audit and EU data residency options. Best where the process is complex and the compliance bar is high.
We build: agent topics, actions, instructions, data grounding, approval flows, evaluation sets.
Fastest to a working pilot
HubSpot · Breeze
Specialised agents — Customer, Prospecting, Data — inside the Smart CRM, with support across chat, email, WhatsApp and voice. Less configuration than Salesforce, live in weeks rather than quarters, and a good fit for marketing-led teams.
We build: agent instructions, knowledge base, handoff rules, CRM property mapping, workflow triggers.
AI next to the operation
Odoo · AI in the suite
AI embedded in the modules rather than sold as a separate agent platform: lead scoring, enrichment, sentiment, drafted replies and natural-language queries. Self-hosted retrieval keeps sensitive data inside your own database, which matters where residency is non-negotiable.
We build: module configuration, RAG on your own records, custom actions, ERP-side tool access.
Integrations
An agent is only as useful as what it can reach
A common reason a CRM agent gives a useless answer is that the answer lives in another system. These are the connections we build most often, and what moves in each direction.
Every connection is documented with its direction, retry behaviour, failure queue and owner. An integration without a failure path is not finished.
Fit
Where this works, and where it does not
Being explicit about the second list saves both sides a discovery call. If you are in the right-hand column, we will tell you what to fix first — even when that means not buying this.
You have a CRM, a queue and a baseline
- A CRM in daily use for at least six months
- Enough volume to measure: roughly 500+ enquiries a month or 50+ tickets a week
- One named owner for the process being automated
- A team willing to approve outputs for a four-week pilot
- Permission to log and review what the agent did
Not yet, and here is what comes first
- No CRM yet, or a migration in the next quarter — do the migration first
- Duplicate rates high enough that the agent retrieves the wrong account — start with a diagnostic
- Nobody available to approve or correct outputs — the agent will be switched off
- A website widget with no CRM connection — that is a chat tool, not CRM automation
- A decision that must be fully autonomous with no human in the loop
Scope
What is in the engagement, and what is not
Platform AI licences are bought from the vendor, not marked up by us. That keeps the recommendation honest and the invoice readable.
| Item | Inside the engagement | Separate decision |
|---|---|---|
| Audit, agent design and evaluation set | Included | — |
| Agent build, tools and CRM actions | Included | — |
| Approval flows, logging and kill switch | Included | — |
| Pilot measurement and written results | Included | — |
| Platform AI licences and consumption | — | Bought directly from the vendor |
| Data cleanup at scale | — | Diagnostic, then a scoped project |
| Custom model training or fine-tuning | — | Quoted separately, rarely justified |
| Human agents to run escalations | — | Your team, or a support retainer |
| Ongoing monitoring and tuning | — | AI operations retainer |
One workflow, one team, one measurable baseline. Widening scope before the pilot produces a demo, not a deployment.
FAQ
CRM AI and automation, answered
What is CRM AI automation?
Which platform AI should we use: Agentforce, Breeze or Odoo AI?
Do we need clean data before starting?
Will the AI send emails or change records on its own?
How do you measure whether it works?
Is this GDPR-compliant, and what about the EU AI Act?
How long does a deployment take?
What does it cost?
Can you work with our existing CRM partner or internal admin?
Start with one workflow
Find the workflow worth automating before you buy the licence
Bring one process and one queue. We will tell you whether it is ready for an agent, what the data underneath looks like, and what the measured result would have to be to justify it.